Self-Correcting Multi-Agent AI Research System

A multi-agent AI system that uses tools, research agents, and a CRAG-based architecture.

Comment

Inspiration

I wanted to push my AI engineering skills to their limits by building a system that goes beyond traditional chatbots. Most AI assistants fail when given incomplete or irrelevant data.

•So, I built a system that can refine its responses using CRAG, perform deep research, and fetch real-time data using tools and MCP integrations.

What it does

This project is a Self-Correcting Multi-Agent AI Research System that:

  1. Chats with users and persists conversations using Supabase with Short Term memory trigger after every 6 coversation
  2. Uses a CRAG pipeline to detect irrelevant information and rewrite queries
  3. Performs deep research using a dedicated research agent with parallel execution
  4. Integrates multiple tools: • Web search • Calculator • Stock data • Currency conversion • Weather data
  5. Supports image generation during deep research workflows

👉 The system continuously improves its responses by validating and refining retrieved data.

How I built it

  1. Built using LangGraph for multi-agent orchestration
  2. Implemented ReAct agent architecture for reasoning and tool usage
  3. Designed a CRAG (Corrective RAG) pipeline for self-correction

  4. Built a planning + worker-based deep research agent

  5. Tech stack: •OpenAI (LLM reasoning) •Tavily (search) •Supabase (persistent memory) •FAISS (vector retrieval) •Gemini (Image Generation)

  6. Implemented async workflows for parallel research execution.

Challenges we ran into

  1. Deploying a complex multi-agent system on Render free tier with limited resources
  2. Managing asynchronous execution and parallel workflows
  3. Reducing API cost while maintaining CRAG performance
  4. Ensuring secure handling of API keys (in-memory, zero-trust approach)

Accomplishments that we're proud of

  1. Built a working multi-agent AI system with CRAG
  2. Successfully implemented query refinement pipeline
  3. Integrated multiple agents (chat + research) into one system
  4. Achieved working deployment under strict resource constraints

What we learned

  1. Designing multi-agent systems using LangGraph
  2. Handling state, memory, and async execution
  3. Building scalable RAG pipelines with evaluation loops as Subgraph
  4. Integrating tools, MCP and research workflows into a unified system

What's next for Self-Correcting Multi-Agent AI Research System

We plan to build a “Universal Brain” agent, capable of:

  1. Autonomously building and executing projects
  2. Running code securely in sandboxed environments
  3. Coordinating multiple specialized agents for complex tasks

Built With

  • faiss
  • langgraph
  • mcp
  • python
  • render
  • streamlit
  • supabase
Share this project:

Updates